A deep learning based known plaintext attack method for chaotic cryptosystem

Fusen Wang, Jun Sang, Qi Liu, Chunlin Huang, Jinghan Tan

Abstract

In this paper, we propose a known-plaintext attack (KPA) method based on deep learning for traditional chaotic encryption schemes. We employ the convolutional neural network to learn the operation mechanism of chaotic cryptosystem, and accept the trained network as the final decryption system. To evaluate the attack performance of different networks on different chaotic cryptosystems, we adopt two neural networks to perform known-plaintext attacks on two distinct chaotic encryption schemes. The experimental results demonstrate the potential of deep learning-based method for known-plaintext attack against chaotic cryptosystems. Different from the previous known-plaintext attack methods, which were usually limited to a specific chaotic cryptosystem, a neural network can be applied to the cryptanalysis of various chaotic cryptosystems with deep learning-based approach, while several different networks can be designed for the cryptanalysis of chaotic cryptosystems. This paper provides a new idea for the cryptanalysis of chaotic image encryption algorithms.

Introduction

Chaotic cryptosystems are widely utilized for securing digital images due to their high sensitivity to initial conditions and the production of pseudo-random sequences that resemble noise. However, traditional cryptanalysis methods are often "bespoke," meaning an attack designed for one specific chaotic map usually fails if the algorithm is slightly modified. This research explores whether deep learning can provide a more universal attack method by treating decryption as an image-to-image translation task. Specifically, the authors investigate a Known-Plaintext Attack (KPA) scenario, where an attacker uses a set of original images (plaintexts) and their corresponding encrypted versions (ciphertexts) to train a model that learns to reverse the encryption for unseen data.

Methodology

The core methodology treats decryption as a restoration task, similar to denoising or super-resolution. The study evaluates two primary convolutional neural network (CNN) architectures: a modified U-Net and a novel Multi-Stage Encoder-Decoder Network (MSEDNet).

The U-Net architecture used for the known-plaintext attack, featuring an encoder-decoder structure with skip connections to preserve spatial information.

The U-Net architecture (shown above) utilizes an encoder-decoder structure to capture high-level features and reconstruct fine-grained details. The attack was tested on two benchmark datasets:

  • MNIST: Grayscale images of handwritten digits, encrypted with a standard 1D Logistic Map.
  • CIFAR-10: Color images across 10 classes, encrypted with a more complex Hybrid Chaotic Map (combining Logistic, Sine, and Chebyshev maps).

The training objective was to minimize the L1 loss between the network's output and the ground-truth plaintext until the model effectively learned the decryption logic.

Experimental Results

The experimental results demonstrate that deep learning can almost perfectly reverse these chaotic encryptions. Once trained, the neural network itself becomes the "equivalent key," bypassing the need to mathematically derive original chaotic parameters.

Comparative performance of MSEDNet and U-Net over 200 epochs, showing (a) training loss and (b) test correlation coefficient.

Key findings include:

  • High Accuracy: For the Logistic Map (MNIST), the U-Net achieved a 99.15% testing accuracy (measured by correlation). For the Hybrid Chaotic Map (CIFAR-10), it achieved 95.70%.
  • Architecture Performance: As illustrated in the performance curves, the standard U-Net outperformed the proposed MSEDNet in both convergence speed and final accuracy, suggesting that established image restoration architectures are already highly potent tools for cryptanalysis.
  • Generality: Unlike traditional methods, a single neural network architecture can be adapted to attack entirely different encryption schemes, highlighting a significant vulnerability in current chaotic cryptographic standards.

Conclusion

This study confirms that deep learning-based KPAs are a potent threat to chaotic image encryption. By learning the decryption mapping directly from data, these models bypass the mathematical complexity of chaotic maps. While the method requires a substantial library of plaintext-ciphertext pairs for training, its flexibility and high accuracy mark a significant shift in the field of cryptanalysis. Future research will focus on scaling these attacks to even more complex encryption mechanisms and investigating the minimum data requirements for successful training.

Summary updated

Abstract

In this paper, we propose a known-plaintext attack (KPA) method based on deep learning for traditional chaotic encryption scheme. We employ the convolutional neural network to learn the operation mechanism of chaotic cryptosystem, and accept the trained network as the final decryption system. To evaluate the attack performance of different networks on different chaotic cryptosystem, we adopt two neural networks to perform known-plaintext attacks on two distinct chaotic encryption schemes. The experimental results demonstrate the potential of deep learning-based method for known-plaintext attack against chaotic cryptosystem. Different from the previous known-plaintext attack methods, which were usually limited to a specific chaotic cryptosystem, a neural network can be applied to the cryptanalysis of various chaotic cryptosystems with deep learning-based approach, while several different networks can be designed for the cryptanalysis of chaotic cryptosystems. This paper provides a new idea for the cryptanalysis of chaotic image encryption algorithm.
BibTeX
@misc{wang2021deep,
  author = {Fusen Wang and Jun Sang and Qi Liu and Chunlin Huang and Jinghan Tan},
  title = {{A deep learning based known plaintext attack method for chaotic cryptosystem}},
  year = {2021},
  eprint = {2103.05242},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.05242},
}

Loading related content...

Figure